mirror of
https://github.com/hanzoai/open-instruct.git
synced 2026-11-03 02:13:07 +00:00
77 lines
2.9 KiB
YAML
77 lines
2.9 KiB
YAML
model_name_or_path: /model
|
|
model_revision: main
|
|
use_flash_attn: true
|
|
gradient_checkpointing: true
|
|
dataset_mixer:
|
|
#ai2-adapt-dev/hs2_nvidia_new: 1.0
|
|
#valpy/dpo_rewritten_prompts: 1.0
|
|
#valpy/taxonomy_tulu: 1.0
|
|
#allenai/ultrafeedback_binarized_cleaned_train: 1.0
|
|
#ai2-adapt-dev/DaringAnteater-prefs-RM-filter: 1.0
|
|
#ai2-adapt-dev/only_wildchat_aug28_regenerated_llama: 1.0
|
|
# ai2-adapt-dev/ultrafeedback-replication-p0: 1.0
|
|
#ai2-adapt-dev/ultrafeedback-replication-p1: 1.0
|
|
#ai2-adapt-dev/ultrafeedback-replication-p2: 1.0
|
|
#ai2-adapt-dev/ultrafeedback-replication-p3: 1.0
|
|
#ai2-adapt-dev/ultrafeedback-replication-p4: 1.0
|
|
#ai2-adapt-dev/ultrafeedback-replication-p5: 1.0
|
|
#ai2-adapt-dev/ultrafeedback-replication-p6: 1.0
|
|
# ai2-adapt-dev/hh-rlhf-helpful: 20000
|
|
#ai2-adapt-dev/webgpt-binarized: 1.0
|
|
#ai2-adapt-dev/WildChat-prefs-280824: 1.0
|
|
#ai2-adapt-dev/helpsteer2-binarized-mean-aspects: 1.0
|
|
#ai2-adapt-dev/tulu-2/5-prefs-helpsteer: 1.0
|
|
#ai2-adapt-dev/helpsteer2-binarized-nvidia-spec: 1.0
|
|
#ai2-adapt-dev/helpsteer-2-binarized-above-2.0-margin-0.5-ignore-verbosity: 1.0
|
|
#ai2-adapt-dev/helpsteer-2-binarized-above-2.5-ignore-verbosity: 1.0
|
|
# ai2-adapt-dev/UltraInteract_pair_randomlen_Logic: 1.0
|
|
# ai2-adapt-dev/UltraInteract_pair_randomlen_Math_PoT: 1.0
|
|
# ai2-adapt-dev/UltraInteract_pair_randomlen_Math_CoT: 1.0
|
|
# ai2-adapt-dev/UltraInteract_pair_randomlen_Coding: 1.0
|
|
# ai2-adapt-dev/UltraInteract_pair_maxlen_Logic: 1.0
|
|
#ai2-adapt-dev/UltraInteract_pair_maxlen_Math_PoT: 1.0
|
|
# ai2-adapt-dev/UltraInteract_pair_maxlen_Math_CoT: 1.0
|
|
# ai2-adapt-dev/UltraInteract_pair_maxlen_Coding: 1.0
|
|
# ai2-adapt-dev/Skywork-Magpie: 1.0
|
|
#ai2-adapt-dev/nectar_binarized-dedup-ultrafeedback: 1.0
|
|
#ai2-adapt-dev/nectar_binarized-lmsys-chat-1m: 1.0
|
|
#ai2-adapt-dev/nectar_binarized-anthropic-hh: 1.0
|
|
|
|
#LJ's helpsteer regenerations
|
|
# orig Helpsteer2
|
|
ai2-adapt-dev/helpsteer2-uf-pipeline-regen: 1.0
|
|
|
|
#the newer Helpsteer2-Preferences dataset.
|
|
#ai2-adapt-dev/helpsteer2-prefs-uf-pipeline-regen: 1.0
|
|
|
|
#LJ's UF regenerations
|
|
ai2-adapt-dev/tulu3.4-sft-replica-50k-gpt4-prefs-on-policy: 1.0
|
|
#ai2-adapt-dev/tulu3.4-sft-replica-50k: 1.0
|
|
|
|
#LJ's regenerations of my datasets
|
|
#ai2-adapt-dev/DaringAnteater-prefs-RM-filter-uf-pipeline-regen: 1.0
|
|
#ai2-adapt-dev/WildChat-prefs-280824-uf-pipeline-regen: 1.0
|
|
#ai2-adapt-dev/Llama-3.1-if_taxonomy_tulu-uf-regen: 1.0
|
|
|
|
#Faeze's persona IF preference dataset
|
|
ai2-adapt-dev/personahub_if_pref_data_manualseed_v2_19890: 1.0
|
|
tokenizer_name: /model
|
|
use_slow_tokenizer: true
|
|
max_seq_length: 2048
|
|
preprocessing_num_workers: 16
|
|
per_device_train_batch_size: 1
|
|
gradient_accumulation_steps: 16 # designed for 8 GPUs, so batch size 128
|
|
learning_rate: 5.0e-7
|
|
lr_scheduler_type: linear
|
|
warmup_ratio: 0.1
|
|
weight_decay: 0.0
|
|
num_train_epochs: 1
|
|
output_dir: /output
|
|
with_tracking: true
|
|
report_to:
|
|
- wandb
|
|
logging_steps: 1
|
|
use_lora: false
|
|
dpo_loss_type: dpo_norm
|
|
dpo_beta: 5
|
|
checkpointing_steps: 1000 |